How does this fit into the mainstream AI/ML pipeline which generally happens in Python ecosystem?

[khteh]

Having the benefits of speed and performance is great but what's next? How to use the data in the mainstream AI/ML pipeline which generally happens in Python ecosystem in order to train and build a model using the mainstream framework like Tensorflow?

[hosseinmoein]

I am not sure what you mean. Your question is too general to answer.
Do you have a specific example/problem that you can describe in some detail?

In general it should be easy to pick up data at rest or even data in motion using the C++ DataFrame and process it. It doesn't matter what is on the other side.

The problem that I hear often is that:

1. C++ is too difficult to learn and master
2. Because of the compile/link cycle in C++, it takes longer to develop ideas.

I will be happy to discuss both of them and argue that both are untrue when considering the big picture.

[khteh]

For example,
(1) How to use this in Jupyter Notebook, especially for eager execution?
(2) How to interface and work with Tensorflow with Keras for NN? I find some references in SO but most, if not all, of them are building the model in Python, then covert and use them in C++ application. However, if the data manipulation result of this library cannot be directly and efficiently used to build a model, like what we are used to in Jupyter Notebook or a single python script workflow, I feel that there is a gap / disconnect here.

I come from C/C++/C# (ML.Net) and compilation/linking is natural for me. If one has problem with this, he/she will have problem with other similar programming languages like Java and should not choose this language for their solution.

[hosseinmoein]

For Jupyter Notebook, you can use Xeus or https://cpp.sh/, or ...

Tensorflow itself is implemented in C++. Its primary interface is C++ interface. The Python interface is just bindings on top of the C++ interface. Indeed, it is more natural (you can do more) to use Tensorflow with C++ than Python.
The fact that most models are implemented in Python is a choice not necessity. You would have more flexibility if you do it in C++. Again IMO, the reason that Python is more popular is because of the two ill-perceived issues I mentioned above.